Abstract: In response to the problems of significant domain difference in input images, blurred detection objects, target occlusion, low number of target samples, and poor quality pseudo labels in foggy scene target detection models, an unsupervised cross-domain target detection model called UMT-HPLA (UMT-Harmony Pseudo Label Assigner) based on Unbiased Mean Teacher (UMT) is proposed. The model consists of a non-paired image generator, a YOLOv5 detector, and a pseudo label processing module, which are improved from the perspectives of domain adaptation, detection performance optimization, and pseudo label quality enhancement. The non-paired image generator generates style-consistent pseudo images by contrastive learning to effectively alleviate the domain difference problem. The YOLOv5 detector introduces the Varifocal Loss (VFL) to enhance the robustness to occluded targets. The pseudo label processing module includes a pseudo label selector and a pseudo label assigner. The pseudo label selector selects samples by combining the location and confidence as the dominant factors to provide high-quality pseudo labels, and the new designed pseudo label assignment method adopts a differentiated loss calculation strategy for different types of pseudo labels.The experimental results on the public datasets Cityscapes and Foggy Cityscapes show that compared with the UMT method with YOLO as the detector, the mAP of the proposed model is increased by 3.7%, reaching 58.1%.